import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a HardSigmoid activation.
    """
    def __init__(self):
        super(Model, self).__init__()
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies HardSigmoid activation to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of any shape.

        Returns:
            torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
        """
        return torch.nn.functional.hardsigmoid(x)

batch_size = 4096
dim = 393216

def get_inputs():
    x = torch.rand(batch_size, dim)
    return [x]

def get_init_inputs():
    return []  # No special initialization inputs needed

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>

__device__ __forceinline__ float hs_f32(float a) {
    // PyTorch hardsigmoid is clamp(a + 3, 0, 6) / 6
    a = fminf(fmaxf(a + 3.0f, 0.0f), 6.0f);
    return a * 0.16666666666666666f;
}

__global__ void hardsigmoid_vec4_kernel(const float* __restrict__ x,
                                        float* __restrict__ y,
                                        long long n4) {
    long long i = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    long long stride = (long long)blockDim.x * gridDim.x;
    const float4* __restrict__ x4 = reinterpret_cast<const float4*>(x);
    float4* __restrict__ y4 = reinterpret_cast<float4*>(y);
    for (; i < n4; i += stride) {
        float4 v = x4[i];
        v.x = hs_f32(v.x);
        v.y = hs_f32(v.y);
        v.z = hs_f32(v.z);
        v.w = hs_f32(v.w);
        y4[i] = v;
    }
}

__global__ void hardsigmoid_tail_kernel(const float* __restrict__ x,
                                        float* __restrict__ y,
                                        long long start,
                                        long long n) {
    long long i = start + (long long)blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n) y[i] = hs_f32(x[i]);
}

__global__ void hardsigmoid_scalar_kernel(const float* __restrict__ x,
                                          float* __restrict__ y,
                                          long long n) {
    long long i = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    long long stride = (long long)blockDim.x * gridDim.x;
    for (; i < n; i += stride) y[i] = hs_f32(x[i]);
}

torch::Tensor hardsigmoid_forward(torch::Tensor x) {
    auto y = torch::empty_like(x);
    const long long n = x.numel();
    if (n == 0) return y;

    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    const int threads = 256;

    uintptr_t ax = reinterpret_cast<uintptr_t>(xp);
    uintptr_t ay = reinterpret_cast<uintptr_t>(yp);
    if (((ax | ay) & 15ULL) == 0ULL && n >= 4) {
        long long n4 = n >> 2;
        int sms = at::cuda::getCurrentDeviceProperties()->multiProcessorCount;
        int blocks = (int)((n4 + threads - 1) / threads);
        int max_blocks = sms * 8;
        if (blocks > max_blocks) blocks = max_blocks;
        if (blocks < 1) blocks = 1;
        hardsigmoid_vec4_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n4);
        long long done = n4 << 2;
        if (done < n) {
            hardsigmoid_tail_kernel<<<1, threads, 0, stream>>>(xp, yp, done, n);
        }
    } else {
        int sms = at::cuda::getCurrentDeviceProperties()->multiProcessorCount;
        int blocks = (int)((n + threads - 1) / threads);
        int max_blocks = sms * 8;
        if (blocks > max_blocks) blocks = max_blocks;
        if (blocks < 1) blocks = 1;
        hardsigmoid_scalar_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n);
    }
    return y;
}
'''

_ext = load_inline(
    name='hardsigmoid_inline_ext_fast',
    cpp_sources='',
    cuda_sources=_cuda_src,
    functions=['hardsigmoid_forward'],
    with_cuda=True,
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    extra_cflags=['-O3'],
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self):
        super(ModelNew, self).__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.hardsigmoid_forward(x)